With the development of artificial intelligence (AI) technology, the autonomous navigation and behavior decision-making capabilities of MASS (marine autonomous surface ship) are constantly being innovated, thereby ensuring their safe navigation. However, the recent algorithms exhibit limited efficacy in navigating in unknown and complex environments, while also lacking the capability to effectively handle the encounters resulting from the uncertain behavior of other ships. Consequently, this study proposes an intelligent navigation methodology utilizing the PRM (Probabilistic Roadmap) and PPO (Proximal Policy Optimization) algorithm to facilitate autonomous navigation and collision avoidance decision-making for MASS. Moreover, the MASS disciplined behaviors prescribed by COLREGs are taken into the consideration of the reward function design. Particularly, in extreme encounter situation, it becomes necessary for MASS to depart from COLREGs, thus requiring a corresponding definition of the reward function. Finally, the autonomous navigation and decision-making capability of the MASS is evaluated using real-time ship traffic in a voyage scenario, while various extreme encounter situations are also simulated to demonstrate the generality and practicality of the proposed PRM-PPO method.

随着人工智能（AI）技术的发展，MASS（海洋自主水面舰艇）的自主导航和行为决策能力不断创新，从而保证其安全航行。然而，最近的算法在未知和复杂环境中的导航效果有限，同时也缺乏有效处理由其他船舶的不确定行为引起的遭遇的能力。因此，本研究提出了一种利用 PRM（概率路线图）和 PPO（邻近策略优化）算法的智能导航方法，以促进 MASS 的自主导航和防撞决策。此外，COLREG 规定的 MASS 纪律行为也被纳入奖励函数设计的考虑范围。特别是在极端遭遇情况下，MASS有必要偏离COLREG，从而需要相应的奖励函数定义。最后，利用航行场景中的实时船舶交通评估了MASS的自主导航和决策能力，同时还模拟了各种极端遭遇情况，以证明所提出的PRM-PPO方法的通用性和实用性。